[Feature] Add pointcloud ppo script

[rrzhang139]

Added baseline ppo script for pointclouds. Useful for oracle (teacher-student) policies.

Note: I wasn't able to fully run this script yet due to OOM errors, please let me know if there are any bugs.

[StoneT2000]

One issue is for a true point cloud robot learning baseline we need to add code to support sampling configurations if obs_mode="pointcloud", or to use obs_mode="sensor_data" and parse into pointcloud data manually and then sample. Pointcloud data uses a ton of memory relative to the raw sensor data and typically we only sample some points.

[rrzhang139]

Are there any libraries or code I can reference for adding this sampling configuration? I was thinking of something like passing them directly into the env.

env = gym.make("PickCube-v1", obs_mode="pointcloud", point_cloud_sample=dict(method="fps", n_points=4096))

[rrzhang139]

@StoneT2000 Could you take a look at this new commit?

[StoneT2000]

What are the results like for training with PCD on some tasks like PickCube or StackCube?